AMD's move toward Taalas-style hardwired AI inference silicon challenges the GPU's dominance. Here's why custom chips baking models directly into silicon could reshape the economics of running generative AI — including video and voice models.
A new benchmark systematically tests whether large language models comply with China's AI-generated content regulations, revealing gaps in how models handle labeling, safety, and legal requirements for synthetic media.
New research reveals that text watermarking schemes designed to identify AI-generated content can inadvertently make language models more susceptible to adversarial prompt attacks, exposing a critical tension between authenticity tooling and model security.
Model collapse is the slow-motion degradation that happens when AI models train on AI-generated data. As synthetic media floods the web, future models risk losing diversity, accuracy, and quality in a compounding feedback loop.
As AI video models like Sora, Veo, and Runway close the quality gap, the true production bottleneck has shifted to orchestration — stitching together generation, editing, audio, and consistency into coherent pipelines that scale.
NVIDIA introduces NeMo Data Designer, an extensible framework for generating multimodal synthetic data at scale — a foundation for training the next generation of AI video, image, and audio models.
Nunchux AI's VC-Attention is a training-free, low-bit attention kernel that accelerates video diffusion transformers without retraining—slashing inference costs for AI video generation while preserving output quality.
New research shows how small video-language models can gain advanced reasoning through synthetic chain-of-thought distillation and difficulty-aware fine-tuning, closing the gap with large models at a fraction of the compute cost.
Open-weight Chinese models from labs like DeepSeek and Alibaba are rapidly narrowing the performance gap with Silicon Valley's closed frontier systems—reshaping the economics of AI and the tools that power synthetic media generation.
Sakana AI researchers unveil PC-ALM, a layer-local training method that replaces global backpropagation and successfully trains 1000-layer networks, hinting at more scalable and parallelizable deep learning.
Deepfake detection systems have seen accuracy fall by 49% over the past year as AI video generation tools grow more sophisticated, widening the gap between synthetic media creation and the tools built to catch it.
Speculative decoding lets AI models predict multiple tokens ahead using a small draft model verified by a larger one, dramatically cutting inference latency without sacrificing output quality — a technique reshaping how fast generative models run.